| As time goes to 2020 s,with the development of oil and gas resources exploration,the construction of easy-to-exploit oil and gas mines has been completed.Therefore,the research topic of seismic exploration is gradually shifting to the complex structural and deep areas.To breakthrough the challenges of high-resolution exploration,we not only need to develop the classic seismic exploration technologies but also need to find new exploration technologies.Among them,deep learning in artificial intelligence is undoubtedly outstanding.In recent years,the rapid development of theoretical methods,supporting software and hardware for deep learning has been applied in many fields with great success.Although the application of deep learning in seismic exploration is late,new achievements and breakthroughs are endless,which makes the deep learning research a hot topic.The traditional seismic exploration of acquisition,processing and interpretation,which all need personal experience and high-intensity repetitive operations.In this paper,we incorporated the deep learning technologies into seismic exploration which can not only reduce the working pressure of staff in areas where traditional seismic exploration is mature,but also make improvements in areas where traditional seismic exploration has not yet broken through.In this case,the seismic exploration can make steps to deeper and more complex areas.We applied the deep learning theory to design four convolutional neural networks based on the characteristics of seismic exploration.They are named waste traces detection convolutional neural network,denoising and interpolation convolutional neural network,seismic signal broadband convolutional neural network,and velocity modeling convolutional neural network.According to the names,it is easy to understand that their role is to detect waste traces in seismic data,remove noise from the seismic data and interpolate for waste traces,extend the frequency band of effective seismic signals to both ends and establish the medium velocity model of the exploration area.The corresponding research results of the four convolutional neural networks can be summarized as follows:(1)Waste traces detection convolutional neural network: the waste traces in the seismic data which are caused by the interference of many natural environments and some human factors.The causes of waste traces are roughly divided into four types: empty traces,anti-traces,single frequency interference traces and strong noise interference traces.This paper uses deep learning principles to construct a convolutional neural network to automatically identify waste traces to improve the collection efficiency and reduce the judgment burden of field workers.The function of the designed waste traces detection neural network is to make two judgments on the seismic records: is it a waste trace or what kind of waste trace is it? The special classification labels and cost functions are designed to accomplish this purpose,which have proven to be completely reliable after training,learning and testing on simulated data.Finally,after successfully passing the test of actual seismic data,it proved that the excellent design of the waste traces detection convolutional neural network is a promissing approach to be applied in the future.(2)Denoising and interpolation convolutional neural network: The incomplete seismic data which are caused by random noise and waste traces.To solve this problem,a convolutional neural network is designed to achieve denoising and interpolating.The seismic data in one step compared to the traditional two modules of denoising and interpolating,the processing efficiency has been significantly improved.To achieve the denoising and interpolation purpose of seismic data,the DNCNN neural network is applied,which has the excellent ability to learn random noise.Combining with the concept of a denoising autoencoder,an hourglass fully connected layer is added in the middle of the fully connected convolutional neural network structure.The entire CNN is divided into two parts,one is coding area and the other is decoding area.The input information is first compressed and then enlarged,which can gradually refine the effective label.And the CNN numerical tests of the simulated data which successfully proved that the CNN is a good method for seismic data denoising.The comparison of the final actual test proves that the training set is an important factor for CNN,and also proves a famous saying: the future is in the hands of people who have more data.(3)The CNN based low frequency reconstruction: due to the lack of reliable lowfrequency components of seismic data,the FWI misfit is often trapped in a local minimum,which leads to the inversion failure.In this paper,by considering deep learning as a similar regression operation,a seismic signal broadband CNN algorithm is constructed.This CNN can complete the wide-band operation of seismic data,so that the effective signal of the seismic data is extended to both low-frequencies and highfrequencies.Where the low-frequencies are learned from the seismic data which are lacking of low-frequency and high-frequency components.After the training and the testing of simulation data,the CNN is applied to the FWI.Comparison of the inversion results of with and without using the CNN methods,the low-frequency components reconstructed by the CNN can effectively establish the macro information of the velocity model,and provide a good start for the application of FWI.And the highfrequency component can make the FWI result clearer and the underground structure more intuitive.Finally,in view of the inability to judge the effect of wideband processing of actual seismic data,a strategy for enlarging the target is proposed.By generalizing the target area,the probability that the training target of the CNN contains the real target is increased.And make the output of seismic signal broadband processing CNN more feasible.(4)The CNN based FWI algorithm: Aiming at the inefficiency caused by the huge amount of FWI,we attempt to use deep learning theory to build a CNN.It is expected that it can directly extract the velocity model from seismic data in order to completely skip the FWI step,and called it velocity modeling CNN.Since velocity modeling can be regarded as an alternative image processing,CNNs are still selected as the overall framework of this CNNs.However,the amount of seismic data is too large,and direct input without optimization will cause a huge burden on the CNN.Therefore,the selfexcitation and self-receiving records of seismic data are finally selected as input.This can ensure that the data carries as much underground information as possible with the minimum amount of data.After training and testing with simulated data,it was found that although the output of the velocity modeling CNN was not obviously wrong,there was still a certain gap with the actual velocity.And if this output velocity model is combined with FWI,excellent high-precision velocity modeling results can also be obtained.Finally,although the velocity modeling CNN constructed in this paper is not yet mature,it has shown great potential in simulation data experiments.It is worth investing more energy in the future and I believe that it will get a satisfactory return. |
PDF Full Text Request